# Direct outputs not eligible for public release

# E.g., to run
# nohup R CMD BATCH --no-save --no-restore '--args outcomes=c("has_ssinc","nonemployed","two_years_nonemployed","five_years_nonemployed")' & #nolint

args <- (commandArgs(TRUE))
if (length(args) > 0) {
    for (i in 1:length(args)) {
        eval(parse(text = args[[i]]))
    }
}

# GLOBAL SETTINGS --------------------------------------------------------------

options(
    scipen = 999,
    digits = 16,
    max.print = .Machine$integer.max,
    show.error.locations = TRUE,
    warn = 1
)

RNGkind("L'Ecuyer-CMRG")
seed <- 818675309L
set.seed(seed) # setting main seed

# PACKAGES ---------------------------------------------------------------------
library(data.table)
library(Matrix)
library(zoo)
library(multiwayvcov)
library(lmtest)
library(checkmate)
library(futile.logger)
library(lfe)

library(remotes)
remotes::install_github("setzler/eventStudy/eventStudy")
library(eventStudy) # for DiD-IV

# PACKAGE SETTINGS -------------------------------------------------------------

# data.table
setDTthreads(threads = 1L)
options(datatable.print.class = TRUE, datatable.print.keys = TRUE)
# so that printing the data.table also shows the variable type on top

# BEGIN FILE -------------------------------------------------------------------

base::source("~/code/0-utility-functions/wald_es.R", local = TRUE)

MakeWealthRetireEffects <- function(outcome) {

    # Read in lottery panel data
    lottery_panel <- readRDS("~/population-panel-data/lottery_panel_data.rds")

    # For all outcomes, replace with zero if missing
    if (
        (class(lottery_panel[[outcome]]) == "integer") ||
            (class(lottery_panel[[outcome]]) == "numeric")
    ) {
        lottery_panel[is.na(get(outcome)), (outcome) := 0]
    }

    # Similarly, for AGI (used to construct quartiles)
    lottery_panel[is.na(per_adult_adjgross), per_adult_adjgross := 0]

    # Constrain to the living, and tag those 62-64
    # - also tag those without prior OASI to that outcome
    lottery_panel <- lottery_panel[dead == 0]
    lottery_panel[, age_case := as.logical(between(age, 62, 64, incbounds = TRUE))]
    if (outcome == "has_ssinc") {
        lottery_panel[, retire_case := as.logical(has_ssinc == 0)]
    } else {
        lottery_panel[, retire_case := TRUE]
    }

    anticipation <- 0
    omitted_event_time <- -2

    # For two-year exit, need to constrain later winners to win 2+ years later
    if (outcome == "two_years_nonemployed"){
        anticipation <- 1
    }

    # For five-year exit, need to constrain later winners to win 5+ years later
    # - but also normalized against -5 (since outcome has 5-year window)
    if (outcome == "two_years_nonemployed"){
        omitted_event_time <- -5
        anticipation <- 4
    }

    # Produce deliberately aggregated estimates
    collapse_table <-
        data.table(
            a = c("one_to_two", "three_to_five", "post_avg"),
            b = c(list(1:2), list(3:5), list(1:5))
        )

    # Run stacked event-study regression
    did_results_temp <-
        Wald_ES2(
            long_data = copy(lottery_panel),
            outcomevar = outcome,
            unit_var = "tin",
            cal_time_var = "tax_yr",
            onset_time_var = "win_yr",
            cluster_vars = "tin",
            omitted_event_time = omitted_event_time,
            discrete_covars = "age",
            control_subset_var = "age_case",
            control_subset_event_time = 0,
            treated_subset_var = "age_case",
            treated_subset_event_time = 0,
            control_subset_var2 = "retire_case",
            control_subset_event_time2 = omitted_event_time,
            treated_subset_var2 = "retire_case",
            treated_subset_event_time2 = omitted_event_time,
            heterogeneous_only = TRUE,
            anticipation = anticipation,
            endog_var = "L_multiperiod",
            calculate_collapse_estimates = TRUE,
            collapse_inputs = collapse_table,
            add_unit_fes = TRUE
        )

    did_results_q1_temp <-
        Wald_ES2(
            long_data = copy(lottery_panel),
            outcomevar = outcome,
            unit_var = "tin",
            cal_time_var = "tax_yr",
            onset_time_var = "win_yr",
            cluster_vars = "tin",
            omitted_event_time = omitted_event_time,
            discrete_covars = "age",
            control_subset_var = "age_case",
            control_subset_event_time = 0,
            treated_subset_var = "age_case",
            treated_subset_event_time = 0,
            control_subset_var2 = "retire_case",
            control_subset_event_time2 = omitted_event_time,
            treated_subset_var2 = "retire_case",
            treated_subset_event_time2 = omitted_event_time,
            heterogeneous_only = TRUE,
            anticipation = anticipation,
            endog_var = "L_multiperiod",
            calculate_collapse_estimates = TRUE,
            collapse_inputs = collapse_table,
            add_unit_fes = TRUE,
            ntile_var = "per_adult_adjgross",
            ntile_event_time = omitted_event_time,
            ntiles = 4,
            ntile_var_value = 1,
            ntile_avg = FALSE
        )

    did_results_q4_temp <-
        Wald_ES2(
            long_data = copy(lottery_panel),
            outcomevar = outcome,
            unit_var = "tin",
            cal_time_var = "tax_yr",
            onset_time_var = "win_yr",
            cluster_vars = "tin",
            omitted_event_time = omitted_event_time,
            discrete_covars = "age",
            control_subset_var = "age_case",
            control_subset_event_time = 0,
            treated_subset_var = "age_case",
            treated_subset_event_time = 0,
            control_subset_var2 = "retire_case",
            control_subset_event_time2 = omitted_event_time,
            treated_subset_var2 = "retire_case",
            treated_subset_event_time2 = omitted_event_time,
            heterogeneous_only = TRUE,
            anticipation = anticipation,
            endog_var = "L_multiperiod",
            calculate_collapse_estimates = TRUE,
            collapse_inputs = collapse_table,
            add_unit_fes = TRUE,
            ntile_var = "per_adult_adjgross",
            ntile_event_time = omitted_event_time,
            ntiles = 4,
            ntile_var_value = 4,
            ntile_avg = FALSE
        )

    # Above results contain more than is needed for this step, so focus them
    # with a quick subsetting function

    MakeMainRetirementEstimates <- function(did_dt) {

        # reduced-form / event-study estimates
        doutcome_dt <-
            setDT(did_dt[[1]])[
                rn == "att" &
                    model == "reduced_form" &
                    ref_onset_time == "Cohort-Weighted",
                .(ref_event_time, estimate, cluster_se, model)
            ]

        # post-period avg collapsed estimate
        doutcome_dl_dt_collapsed <-
            setDT(did_dt[[1]])[
                rn == "att" &
                    model == "ratio" &
                    ref_onset_time == "Cohort-Weighted + Collapsed" &
                    grouping %in% collapse_table$a,
                .(ref_event_time, estimate, cluster_se, grouping, model)
            ]

        # Will use a 97, 98, 99 convention to get
        # the right order of results in tables
        doutcome_dl_dt_collapsed[
            grouping %in% c("one_to_two", "first_two"),
            ref_event_time := 97
        ]
        doutcome_dl_dt_collapsed[
            grouping == "three_to_five",
            ref_event_time := 98
        ]
        doutcome_dl_dt_collapsed[grouping == "post_avg", ref_event_time := 99]
        doutcome_dl_dt_collapsed[, grouping := NULL]

        # Counterfactual untreated mean of the treated group
        # if E[y(1)-y(0)|D=1] = E[y1-y0|D=1] - E[y1-y0|D=0] (= Reduced form DiD)
        # then E[y(0)|D = 1] = E[y(1)] - RF DiD
        did_cohort_et <-
            setDT(did_dt[[1]])[
                model == "reduced_form" & rn == "catt",
                .(
                    ref_onset_time,
                    ref_event_time,
                    estimate,
                    catt_treated_unique_units
                )
            ]
        setnames(did_cohort_et, "estimate", "did")

        post_mean <-
            setDT(did_dt[[1]])[
                model == "reduced_form" &
                    rn == "treatment_means" &
                    treated == 1,
                .(
                    ref_onset_time,
                    ref_event_time,
                    estimate
                )
            ]
        setnames(post_mean, "estimate", "post_mean")

        cfactual_et <-
            merge(
                did_cohort_et,
                post_mean,
                by = c("ref_onset_time", "ref_event_time"),
                all.x = TRUE,
                sort = FALSE
            )
        # note: of course, omitted_event_time isn't in the above

        post_mean <- NULL
        did_cohort_et <- NULL
        rm(post_mean, did_cohort_et)

        cfactual_et[, estimate := (post_mean - did)]
        cfactual_et[, c("post_mean", "did") := NULL]

        # Will also want cohort-weighted event-time versions
        cfactual_et[, ref_onset_time := as.character(ref_onset_time)]

        cfactual_t <-
            cfactual_et[
                ,
                .(
                    ref_onset_time = "Cohort-Weighted",
                    estimate =
                        weighted.mean(
                            x = estimate,
                            w = catt_treated_unique_units
                        )
                ),
                by = .(ref_event_time)
            ]

        cfactual_collapsed <-
            cfactual_t[
                ref_event_time %in% c(1:5),
                .(
                    ref_onset_time = "Cohort-Weighted + Collapsed",
                    ref_event_time = 99L,
                    estimate = mean(x = estimate)
                )
            ]
        cfactual_collapsed[, cluster_se := 0]
        cfactual_collapsed[, model := "counterfactual_mean"]

        combined <-
            rbindlist(
                list(doutcome_dt, doutcome_dl_dt_collapsed, cfactual_collapsed),
                use.names = TRUE
            )

        return(combined)
    }

    did_results_agg <-
        MakeMainRetirementEstimates(did_dt = did_results_temp)
    did_results_agg[, income_quartile := 5L]

    did_results_q1 <-
        MakeMainRetirementEstimates(did_dt = did_results_q1_temp)
    did_results_q1[, income_quartile := 1L]

    did_results_q4 <-
        MakeMainRetirementEstimates(did_dt = did_results_q4_temp)
    did_results_q4[, income_quartile := 4L]

    did_results_temp <-
        rbindlist(
            list(
                did_results_agg,
                did_results_q1,
                did_results_q4
            ),
            use.names = TRUE
        )

    # Only need the reduced form for the aggregate estimates
    did_results <-
        did_results[
            (model == "reduced_form" & income_quartile == 5) |
                (model != "reduced_form")
        ]

    saveRDS(
        did_results,
        sprintf("~/estimation-output/wealth_effect_estimates_%s_62_to_64.rds", outcome) # nolint
    )

    return(outcome)
}

num_cores <- length(outcomes)
mcmapply(
    MakeWealthRetireEffects,
    outcomes,
    SIMPLIFY = FALSE,
    mc.silent = FALSE,
    mc.cores = num_cores,
    mc.set.seed = TRUE
)
